Method and system for evaluating health of ancient trees

Abstract

The application discloses a kind of ancient tree health assessment method and system, wherein the method comprises: collecting the multidimensional health index data of target ancient tree;The index data is cleaned, normalized and treated with missing values to form a standardized dataset, and a feature vector is extracted;Input into the pre-trained gradient boosting decision tree model, and output the health grade prediction result of target ancient tree;The model takes minimizing the objective function as the training target, and iteratively trains;The objective function is composed of a loss function and a regularization term, including: weighted cross-entropy loss term, time smoothing regularization term;The system is used to implement the method.The application changes the health assessment of ancient trees from qualitative judgment relying on subjective experience to objective diagnosis based on multidimensional quantitative data and algorithm, significantly improving the scientificity of the evaluation result and the decision support value.

Description

Technical Field

[0001] This invention relates to the field of protection and monitoring of ancient and famous trees, specifically to a method and system for health assessment of ancient and famous trees. Background Technology

[0002] Ancient and famous trees, as precious natural and cultural heritage, possess extremely high ecological, historical, scientific, and aesthetic value. Accurate and timely assessment of their health status is fundamental to implementing scientific protection and rejuvenation. With lifespans spanning hundreds or even thousands of years, the health of ancient and famous trees is the result of complex interactions among multiple dimensions of factors, including above-ground morphology, underground root systems, physiological functions, and the surrounding soil environment. This makes health assessment a typical multivariate, nonlinear, and highly complex technical challenge.

[0003] Existing methods for assessing the health of ancient and famous trees have the following three main technical limitations:

[0004] 1. The subjectivity and one-sidedness of traditional manual evaluation methods:

[0005] Currently, the assessment methods widely used at the grassroots level mainly rely on the on-site visual inspection and experience judgment of forestry experts. Although such methods are direct, they have significant drawbacks: (1) strong subjectivity: the assessment results are easily affected by the personal experience and subjective judgment of the assessors. Different assessors may have huge differences in their judgment of the health level of the same ancient tree, and there is a lack of objective and unified quantitative standards; (2) one-sidedness: the assessment focuses on the easily observable above-ground parts (such as withered branches and leaves, hollow trunks, etc.), while it is difficult to effectively diagnose the key hidden factors that determine the vitality of trees (such as root distribution, soil physicochemical properties, trunk sap flow and other physiological activities), and cannot systematically reveal the root cause of health problems; (3) low efficiency: it is time-consuming and labor-intensive, and it is difficult to meet the needs of large-scale and large-scale ancient tree health surveys and dynamic monitoring.

[0006] 2. The superficiality and data silo problem in the application of modern sensing technology:

[0007] With the development of technology, some studies have begun to try to introduce various sensors (such as multispectral cameras, soil sensors, etc.) for data collection. However, the application of such technologies often remains at the level of simple data listing and single-index threshold alarm. The main problems are: (1) Insufficient fusion of multi-source data: The collected multi-dimensional index data such as aboveground morphology, root structure, soil environment, and physiological function are fragmented and lack an effective fusion analysis model, making it impossible to form a correlation diagnosis between "apparent symptoms-internal physiology-environmental stress"; (2) Lack of assessment model: Data is available, but there is a lack of intelligent assessment models that can digest these multi-dimensional and heterogeneous data and output a comprehensive and quantifiable health level, resulting in a disconnect between technical means and final decision-making.

[0008] 3. The "incompatibility" of directly applying general machine learning models:

[0009] To address the problem of assessment models, existing research has attempted to apply machine learning algorithms (such as decision trees and support vector machines). However, when these general models are applied directly, the unique properties of ancient tree health data are seriously ignored, resulting in poor assessment results. Specifically, this manifests as follows: (1) Insensitivity to extreme data imbalance: The natural distribution of ancient tree health status is extremely uneven, with "healthy" and "sub-healthy" samples accounting for the vast majority, while the key "endangered" samples are extremely rare. The loss function of the general model tends to favor the majority class, resulting in a high rate of missed detection in the prediction of the most critical endangered ancient trees; (2) Lack of time-series dynamic perception: The health evolution of ancient trees is a slow and continuous process, and the annual monitoring data has strong time correlation. The general model treats the data of each year as independent samples, and cannot use historical information to capture the "trend" of health decline. It can only make static judgments, resulting in delayed early warnings; (3) Ignoring the heterogeneity of feature importance: Under different tree species (such as acid-loving and alkaline-loving) and different habitats (urban and mountain forests), the weight of the same indicator (such as soil pH) on health is completely different. The general model attempts to use a single global weight to address all situations, which leads to a decrease in evaluation accuracy and makes it impossible to achieve precise evaluation based on the specific circumstances of each tree.

[0010] In summary, current technologies have yet to provide an objective, accurate, and systematic intelligent assessment method that can deeply integrate multi-dimensional health indicators and specifically address the unique characteristics of ancient tree health data (such as imbalance, temporal sequence, and heterogeneity). This has become a key technological bottleneck restricting the improvement of the scientific protection level of ancient and famous trees. Summary of the Invention

[0011] The technical problem to be solved by the present invention is to provide a method and system for health assessment of ancient and famous trees, which transforms the health assessment of ancient and famous trees from qualitative judgment based on subjective experience to objective diagnosis based on multi-dimensional quantitative data and algorithms, thereby significantly improving the scientific nature and decision support value of the assessment results.

[0012] To solve the above-mentioned technical problems, the technical solution adopted by the method of the present invention includes the following steps:

[0013] S1. Collect multi-dimensional health indicator data of the target ancient and famous trees, including above-ground morphological indicators, underground structure and soil environment indicators, and physiological function indicators;

[0014] S2. Clean, normalize, and handle missing values ​​of the indicator data to form a standardized dataset, and extract feature vectors from the standardized dataset.

[0015] S3. Input the feature vector into a pre-trained gradient boosting decision tree model to obtain the health level prediction result of the target ancient and famous trees;

[0016] The gradient boosting decision tree model is trained in the following way:

[0017] Using a dataset of historical ancient and famous trees, the gradient boosting decision tree model is iteratively trained with the objective function minimized as the training objective.

[0018] The objective function consists of a loss function and a regularization term, including:

[0019] (1) Weighted cross-entropy loss term, which introduces adaptive weights for different health level categories. These weights are related to the inverse of the number of samples in each category, in order to increase attention to categories with few samples.

[0020] (2) Temporal smoothing regularization term, used to penalize the drastic fluctuations in the health score results of the same ancient tree at adjacent time points, so that the health evolution trajectory output by the model conforms to the slow and gradual change characteristics of the health of ancient trees.

[0021] As a further improvement to the method of this invention, the specific definition of the weighted cross-entropy loss term is as follows:

[0022]

[0023] in, It is a sample i The true label, The model predicts the sample. i Category c The probability, For each category The dynamic weight is calculated using the following formula:

[0024]

[0025] in, N The total number of samples, Nc For category c The number of samples, This is a smoothing factor.

[0026] As a further improvement to the method of this invention, the time-series smoothing regularization term is defined as follows:

[0027]

[0028] in M For the number of ancient trees, The time series regularization intensity parameter, , It is the health score of the i-th ancient tree calculated by the model in year t and year t-1.

[0029] As a further improvement to the method of this invention, before model training, the training dataset is subjected to hierarchical oversampling. Specifically, the ancient trees are first divided into layers according to their species and age. For layers with few samples, oversampling is performed using boundary sample synthesis technology based on the nearest neighbor distribution of the samples in the feature space, so as to further improve the model's ability to learn from rare class samples.

[0030] As a further improvement to the method of this invention, during the training process of the gradient boosting decision tree model, when generating each decision tree, dynamic weights are assigned to different health indicator features based on the tree species metadata of the sample ancient trees, and split nodes are selected based on the weighted information gain to adapt to the differences in the sensitivity of different tree species to health indicators.

[0031] As a further improvement to the method of this invention, the gradient boosting decision tree model employs a grouping ensemble strategy for training and prediction, including:

[0032] Training phase: The training data is divided into multiple subsets according to the tree species or site type of the ancient trees, and a sub-gradient boosting decision tree model is trained for each subset.

[0033] Prediction phase: For the target ancient tree to be evaluated, it is first divided into the corresponding group according to its metadata. One or more sub-models of the corresponding group are used to predict the target ancient tree. Then, a meta-classifier is used to combine the prediction results of each sub-model to obtain the final health level.

[0034] As a further improvement to the method of this invention, the gradient boosting decision tree model also adopts a multimodal fusion strategy during training. Specifically, health indicators with the same measurement source are divided into the same data modality; when the decision tree grows, node splitting is performed preferentially within the features of the same data modality; and the representations of each data modality are weighted and fused, with the weights dynamically learned by the attention mechanism.

[0035] As a further improvement to the method of this invention, the index data include: normal leaf rate, multispectral index NDVI, degree of branch dieback, degree of trunk dieback, trunk tilt, degree of bark damage, internal rot of the trunk, trunk sap flow intensity, root health, soil coverage, degree of hardening under the crown, soil pH, soil temperature, soil moisture, soil nutrient content, duration of surface water accumulation, and pest and disease damage level.

[0036] As a further improvement to the method of this invention, in step S2, the extraction of feature vectors includes: for the index data with continuous time series data, constructing its time series derived features, the time series derived features include: the linear trend slope, coefficient of variation and the difference between adjacent time points calculated based on the sliding time window.

[0037] This invention also provides a health assessment system for ancient and famous trees, used to implement the above-mentioned health assessment method for ancient and famous trees, comprising:

[0038] The data acquisition module is configured to collect multi-dimensional health indicator data of the target ancient and famous trees;

[0039] The preprocessing and feature engineering module is configured to preprocess data and extract feature vectors.

[0040] The model prediction module contains a pre-trained gradient boosting decision tree model, which receives feature vectors and outputs health level prediction results.

[0041] The results output module is configured to display and store the health level prediction results.

[0042] Compared with the prior art, the advantages of the present invention are as follows:

[0043] The health assessment method and system for ancient and famous trees provided by this invention, through the construction of a multi-source quantitative index system and the introduction of a gradient boosting decision tree model deeply optimized for the characteristics of ancient tree data, brings about beneficial effects mainly in the following aspects:

[0044] First, it achieves a fundamental shift in assessment conclusions from "subjective and one-sided" to "objective and systematic." Traditional methods rely on manual visual inspection, resulting in highly subjective and difficult-to-verify results. This invention constructs a comprehensive health profile of ancient and famous trees by covering quantitative indicators of above-ground, underground, physiological, and environmental aspects. More importantly, the model can learn the complex nonlinear relationships between these indicators, enabling correlation diagnosis and thus accurately locating the root cause of health problems. This provides a direct and scientific basis for precise rejuvenation of "one tree, one policy."

[0045] Second, it solves the core problem of the general model being "unsuitable" in the application of ancient trees, and improves the model's perception and decision-making intelligence. This invention does not simply apply existing algorithms, but has carried out a series of special optimizations: (1) By introducing an adaptive category weight loss function and a hierarchical sampling strategy, it effectively overcomes the problem of the extremely uneven distribution of the health status of ancient trees, and greatly improves the sensitivity of identification of key minority categories such as "endangered"; (2) By designing a time-series regularization term and constructing trend features, it enables the model to perceive the "slow evolution" of health, and realizes the upgrade from static assessment to dynamic early warning; (3) By using dynamic feature weights based on tree species and habitat and a group learning mechanism, it enables the model to understand the personalized characteristics of different ancient trees, and realizes the accurate assessment of "one tree, one policy".

[0046] Third, it opens up a technical path from "multi-source data" to "fusion decision-making", enhancing the interpretability and practicality of the model. The multimodal tree structure and hierarchical attention fusion mechanism adopted in this invention can automatically weigh the importance of data from different sources in each evaluation and output interpretable weight analysis. This not only enhances the credibility of the results, but also enables maintenance personnel to quickly understand the logic of the model's decision-making, overcoming the "black box" limitations of traditional machine learning models.

[0047] In summary, this invention successfully upgrades the health management of ancient and famous trees from an experience-based "post-treatment" model to a proactive and refined protection model driven by data intelligence, which involves "early warning, precise diagnosis, and scientific decision-making." This greatly enhances the scientific nature, foresight, and management efficiency of protection efforts. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating an embodiment of the present invention. Detailed Implementation

[0049] The present invention will be further described below with reference to the accompanying drawings and specific preferred embodiments, but this does not limit the scope of protection of the present invention.

[0050] like Figure 1 As shown, the technical solution adopted in this embodiment includes the following steps:

[0051] S1. Collect multi-dimensional health indicator data of the target ancient and famous trees, including above-ground morphological indicators, underground structure and soil environment indicators, and physiological function indicators;

[0052] S2. Clean, normalize, and handle missing values ​​of the indicator data to form a standardized dataset, and extract feature vectors from the standardized dataset.

[0053] S3. Input the feature vector into a pre-trained gradient boosting decision tree model to obtain the health level prediction result of the target ancient and famous trees;

[0054] The gradient boosting decision tree model is trained in the following way:

[0055] Using a dataset of historical ancient and famous trees, the gradient boosting decision tree model is iteratively trained with the objective function minimized as the training objective.

[0056] The objective function consists of a loss function and a regularization term, including:

[0057] (1) Weighted cross-entropy loss term, which introduces adaptive weights for different health level categories. These weights are related to the inverse of the number of samples in each category, in order to increase attention to categories with few samples.

[0058] (2) Temporal smoothing regularization term, used to penalize the drastic fluctuations in the health score results of the same ancient tree at adjacent time points, so that the health evolution trajectory output by the model conforms to the slow and gradual change characteristics of the health of ancient trees.

[0059] This embodiment addresses the characteristics of ancient and famous trees (imbalance, temporality, and heterogeneity). It employs a gradient boosting decision tree model with a customized objective function. By introducing temporal regularization and trend features, the model can capture subtle trends of health decline, achieving a leap from static assessment to dynamic early warning. This shifts the protection of ancient trees from a passive post-event remediation model to an active early intervention model. Furthermore, by constructing a quantitative indicator system covering "physiological function, structural safety, and habitat suitability," and utilizing the optimized model to learn the complex relationships between indicators, it achieves a transformation from subjective experience-based judgment to objective data-driven approaches, enabling precise identification of the root causes of health problems.

[0060] In a specific application example, the weighted cross-entropy loss term is specifically defined as follows:

[0061]

[0062] in, It is a sample i The true label, The model predicts the sample. i Category c The probability, For each category The dynamic weight is calculated using the following formula:

[0063]

[0064] in, N The total number of samples, Nc For categoryc The number of samples, This is a smoothing factor.

[0065] In a specific application example, the time-series smoothing regularization term is defined as follows:

[0066]

[0067] in M For the number of ancient trees, The time series regularization intensity parameter, , It is the health score of the i-th ancient tree calculated by the model in year t and year t-1.

[0068] In specific application examples, before model training, the training dataset is subjected to stratified oversampling. Specifically, the ancient trees are first stratified according to their species and age. For the strata with a small number of samples, the boundary sample synthesis technique is used to oversample based on the nearest neighbor distribution of the samples in the feature space, so as to further improve the model's ability to learn rare class samples.

[0069] In practical implementation, since there are very few samples in the "endangered" category of ancient and famous trees data, they are like a few isolated points in the high-dimensional feature space composed of multidimensional health indicators. In order to balance the number of samples in the dataset and avoid the model ignoring these key categories, this embodiment adopts an improved version of the SMOTE (Synthetic Minority Over-sampling Technique) algorithm, the Borderline-SMOTE algorithm, to complete the oversampling process. After identifying noise samples, boundary samples, and safe samples in the minority class samples, the boundary samples are oversampled. This balances the number of samples in the dataset and avoids generating a large number of interference samples during the oversampling process, thereby improving the effectiveness and robustness of the model in dealing with extreme imbalances in the health data of ancient trees.

[0070] In a specific application example, during the training process of the gradient boosting decision tree model, when generating each decision tree, dynamic weights are assigned to different health indicator features based on the tree species metadata of the sample ancient trees, and split nodes are selected based on the weighted information gain to adapt to the differences in the sensitivity of different tree species to health indicators.

[0071] In a specific application example, the gradient boosting decision tree model employs a grouping ensemble strategy for training and prediction, including:

[0072] Training phase: The training data is divided into multiple subsets according to the tree species or site type of the ancient trees, and a sub-gradient boosting decision tree model is trained for each subset.

[0073] Prediction phase: For the target ancient tree to be evaluated, it is first divided into the corresponding group according to its metadata. One or more sub-models of the corresponding group are used to predict the target ancient tree. Then, a meta-classifier is used to combine the prediction results of each sub-model to obtain the final health level.

[0074] In a specific application example, the gradient boosting decision tree model also employs a multimodal fusion strategy during training. Specifically, health indicators with the same measurement source are divided into the same data modality; during decision tree growth, node splitting is prioritized within the features of the same data modality; and the representations of each data modality are weighted and fused, with the weights dynamically learned by the attention mechanism.

[0075] In this embodiment, health indicators that are measured from the same source are first classified into the same data modality, such as:

[0076] Optical image modalities include normal leaf rate, multispectral index NDVI, etc. Their representation may be a few high-level features obtained by nonlinearly combining these raw percentages and index values ​​through a small neural network submodule, which can be understood as a quantitative representation of "canopy photosynthetic activity".

[0077] Soil environmental modalities include soil pH, soil temperature, soil moisture, and soil nutrient content. These are likely represented by feature vectors extracted from a smaller neural network submodule, reflecting the overall "root zone water, fertilizer, air, and heat status."

[0078] Morphological and structural modalities include the degree of branch and twig damage, trunk damage, bark damage, and internal trunk rot. These can be abstracted as scoring features of the "mechanical integrity of the tree."

[0079] Physiological functional modalities include trunk sap flow intensity, etc. These may be directly related to "tree water transport efficiency".

[0080] In this embodiment, after obtaining the "representation" (i.e., its respective feature vector) of each modality, instead of simply concatenating them and feeding them to the model, an attention mechanism is introduced. This mechanism automatically learns how much weight each modality's "representation" should have under the current prediction task. For example, when evaluating an ancient tree with yellowing leaves, the model may assign higher weights to the representations of optical imagery and soil environment; while when evaluating an ancient tree at risk of tilting, it may pay more attention to the representation of morphological structure.

[0081] Such a multimodal fusion strategy can better uncover the structural information within the data, and through weighted output, it can more clearly show which aspects of data the assessment mainly relied on (e.g., "70% of this assessment is based on canopy conditions and 30% on soil conditions"). Such a model can flexibly respond to ancient trees with different symptoms and achieve more accurate diagnosis.

[0082] In specific application examples, the index data includes: normal leaf rate, multispectral index NDVI, degree of branch dieback, degree of trunk dieback, trunk tilt, degree of bark damage, internal rot of the trunk, trunk sap flow intensity, root health, soil coverage, degree of hardening under the crown, soil pH, soil temperature, soil moisture, soil nutrient content, duration of surface water accumulation, and pest and disease damage level.

[0083] The 17 indicators selected in this embodiment aim to construct a comprehensive diagnostic model covering "physiological function, structural safety, and habitat suitability." These indicators are not simply a collection of factors, but rather based on the principles of dendrology, plant physiology, soil science, and forest pathology. They systematically depict the complete life process of ancient trees, from canopy photosynthetic production, trunk water transport, root absorption and fixation to soil environmental stress, in order to achieve a comprehensive and quantitative assessment of the health status of ancient trees.

[0084] Tables 1 and 2 below show the specific selection criteria and standardized measurement methods for each indicator.

[0085] Table 1 lists some of the indicators designed to assess the growth vitality, structural integrity, and core physiological functions of the visible parts of ancient trees.

[0086]

[0087] Table 2 contains some indicators designed to assess the health of the hidden systems that support the life of ancient trees—the root system and the soil environment.

[0088]

[0089] These 17 indicators constitute a three-dimensional diagnostic network that ranges from appearance to internal factors, from above ground to below ground, and from current status to stress. They are interconnected and corroborate each other: for example, a decrease in trunk sap flow intensity (8) may be related to damaged root health (9) or waterlogging in the soil (16); branch dieback (3) may be the result of pest and disease damage (17) or soil drought (14).

[0090] Through the fusion diagnosis of this set of parameters, the health assessment of ancient and famous trees is no longer based on guesswork of a single symptom, but on a systematic diagnosis based on the fusion of multi-source data. This allows for a more accurate identification of the root cause of health problems, providing a solid scientific basis for the subsequent development of precise protection and rejuvenation plans for each tree.

[0091] Based on the above 17 indicators, the established indicator system and its classification are shown in Table 3:

[0092]

[0093] Based on this indicator system and its classification, the scores of the established health evaluation indicators are shown in Table 4.

[0094]

[0095] Finally, in this embodiment, the description of the health rating level established for ancient and famous trees is shown in Table 5.

[0096]

[0097] In a specific application example, step S2, the extraction of feature vectors includes: constructing time-series derived features for the index data such as soil temperature and humidity, tree trunk sap flow, morphological change indicators, etc., which have continuous time-series data. The time-series derived features include: the linear trend slope, coefficient of variation, and difference between adjacent time points calculated based on a sliding time window.

[0098] In practical applications, the construction of time-series derived features is suitable for dynamic indicators (such as physiological functions, environmental stresses, and slowly developing structural indicators) that can be monitored over long periods by sensors or obtained through periodic inspections. By calculating the changing trends, fluctuations, and abrupt change points of these indicators, the model can perceive the gradual evolution of health, thereby achieving a leap from static assessment to dynamic early warning.

[0099] This embodiment also provides a health assessment system for ancient and famous trees to implement the above-mentioned health assessment method for ancient and famous trees, including:

[0100] The data acquisition module is configured to collect multi-dimensional health indicator data of the target ancient and famous trees;

[0101] The preprocessing and feature engineering module is configured to preprocess data and extract feature vectors.

[0102] The model prediction module contains a pre-trained gradient boosting decision tree model, which receives feature vectors and outputs health level prediction results.

[0103] The results output module is configured to display and store the health level prediction results.

[0104] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0105] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for assessing the health of ancient and famous trees, characterized in that, include: S1. Collect multi-dimensional health indicator data of the target ancient and famous trees, including above-ground morphological indicators, underground structure and soil environment indicators, and physiological function indicators; S2. Clean, normalize, and handle missing values ​​of the indicator data to form a standardized dataset, and extract feature vectors from the standardized dataset. S3. Input the feature vector into a pre-trained gradient boosting decision tree model, and the model outputs the health evaluation index level score of the target ancient and famous trees, thereby obtaining the health level prediction result. The gradient boosting decision tree model is trained in the following way: Using a dataset of historical ancient and famous trees, the gradient boosting decision tree model is iteratively trained with the objective function minimized as the training objective. The gradient boosting decision tree model employs a grouping ensemble strategy for training and prediction, including: Training phase: The training data is divided into multiple subsets according to the tree species or site type of the ancient trees, and a sub-gradient boosting decision tree model is trained for each subset. Prediction phase: For the target ancient tree to be evaluated, it is first divided into the corresponding group according to its metadata. One or more sub-models of the corresponding group are used to predict the target ancient tree. Then, a meta-classifier is used to combine the prediction results of each sub-model to obtain the final health level. During the training process of the gradient boosting decision tree model, a multimodal fusion strategy is also adopted, specifically: health indicators with the same measurement source are divided into the same data modality; when the decision tree grows, node splitting is preferentially performed within the features of the same data modality; and the representations of each data modality are weighted and fused, with the weights dynamically learned by the attention mechanism. The objective function consists of a loss function and a regularization term, including: (1) Weighted cross-entropy loss term, which introduces adaptive weights for different health level categories. These weights are related to the inverse of the number of samples in each category, in order to increase attention to categories with few samples. (2) Temporal smoothing regularization term, used to penalize the drastic fluctuations in the health score results of the same ancient tree at adjacent time points, so that the health evolution trajectory output by the model conforms to the slow and gradual change characteristics of the health of ancient trees.

2. The method for assessing the health of ancient and famous trees according to claim 1, characterized in that, The specific definition of the weighted cross-entropy loss term is as follows: in, It is a sample i The true label, The model predicts the sample. i Category c The probability, For each category The dynamic weight is calculated using the following formula: in, N The total number of samples, Nc For category c The number of samples, This is a smoothing factor.

3. The method for assessing the health of ancient and famous trees according to claim 1, characterized in that, The definition of the time-series smoothing regularization term is: in M For the number of ancient trees, The time series regularization intensity parameter, , It is the health score of the i-th ancient tree calculated by the model in year t and year t-1.

4. The method for assessing the health of ancient and famous trees according to claim 1, characterized in that, Before model training, the training dataset is subjected to stratified oversampling. Specifically, the ancient trees are first stratified according to their species and age. For the strata with few samples, the boundary sample synthesis technique is used to oversample based on the nearest neighbor distribution of the samples in the feature space, so as to further improve the model's ability to learn rare class samples.

5. The method for assessing the health of ancient and famous trees according to claim 1, characterized in that, During the training process of the gradient boosting decision tree model, when generating each decision tree, dynamic weights are assigned to different health indicator features based on the tree species metadata of the sample ancient trees, and split nodes are selected based on the weighted information gain to adapt to the differences in the sensitivity of different tree species to health indicators.

6. The method for assessing the health of ancient and famous trees according to claim 1, characterized in that, The data indicators include: normal leaf rate, multispectral index NDVI, degree of branch dieback, degree of trunk dieback, trunk tilt, degree of bark damage, internal rot of the trunk, trunk sap flow intensity, root health, soil cover, degree of hardening under the canopy, soil pH, soil temperature, soil moisture, soil nutrient content, duration of surface water accumulation, and pest and disease damage level.

7. The method for assessing the health of ancient and famous trees according to claim 1, characterized in that, In step S2, the extraction of feature vectors includes: for the index data with continuous time series data, constructing its time series derived features, the time series derived features include: the linear trend slope, coefficient of variation and the difference between adjacent time points calculated based on the sliding time window.

8. A health assessment system for ancient and famous trees, used to implement the health assessment method for ancient and famous trees as described in any one of claims 1 to 7, characterized in that, include: The data acquisition module is configured to collect multi-dimensional health indicator data of the target ancient and famous trees; The preprocessing and feature engineering module is configured to preprocess data and extract feature vectors. The model prediction module contains a pre-trained gradient boosting decision tree model, which receives feature vectors and outputs health level prediction results. The results output module is configured to display and store the health level prediction results.

Images